Merge alpha: remove top_k workaround + cache key fix

[satra]

HF bug fixed upstream. Pass top_k to pipeline constructor. Remove batch_size from cache key.

[gemini-code-assist]

Code Review

This pull request updates the Hugging Face audio classification pipeline by removing batch_size from the cache key and re-enabling the top_k and function_to_apply parameters in the pipeline constructor. Feedback indicates that including top_k and function_to_apply in the cache key is inefficient and memory-intensive because these parameters are already handled during inference. Additionally, it was noted that the batch_size parameter is currently ignored during the inference call.

[gemini-code-assist]

Including top_k and function_to_apply in the cache key and pipeline constructor is inefficient. These parameters are already passed during the inference call in classify_audios_with_transformers (line 149). By including them in the cache key, a new model instance is loaded into memory for every unique combination of these parameters, which is unnecessary and memory-intensive. Removing them from the key and constructor allows the same pipeline to be reused for different output configurations.

Additionally, the batch_size parameter is now unused in this function and is also missing from the pipeline call at line 149. This means the user-provided batch size is currently ignored during inference, defaulting to 1. You should remove the unused parameters from this function's signature and ensure batch_size is passed to the pipe() call in classify_audios_with_transformers.

        key = f"{model.path_or_uri}-{model.revision}-{device.value}"
        if key not in cls._pipelines:
            cls._pipelines[key] = cast(
                Pipeline,
                pipeline(  # type: ignore[call-overload]
                    task="audio-classification",
                    model=model.path_or_uri,
                    revision=model.revision,
                    device=device.value,
                ),
            )